runtime, the singleton instance organizes and executes tasks.
- Run function that actually drives the execution of the tasks and is thread-safe.
- Queue field that contains the runnable tasks.
taskContainer wraps the logic of operator and maintains the status and input/output channels.
- Poll triggers a executing action of the task, which will return quickly. During the poll action, the task will
  - read inputs channels, if the channels are empty, return blocked. The read action fetches a batch of data.
- status records the current status of task, including
  - runnable means the task is in the Queue and can run immediately.
  - blocked means the task is waiting someone to trigger, say output data consuming, input data arriving or i/o task completed.
  - awaiting means this task has been awoken by someone. If the Poll function ends and checks this status, we should put this task back to queue and reset the status to runnable.
Operator actually implement the user logic. Every operator is contained by a task. Ideally, operators can construct an operating tree like a typical sql engine.
- prepare is called when constructing the operator and prepare some resources.
- close releases all resources gracefully.
- next(ctx *taskContext, r *record, index int) accepts the incoming data and returns the result data or blocked status. index indicates which input the record comes from.
- NextWantedInputIdx returns a integer indicating which input the task should read in this poll. In normal cases it returns non-negative value. There are two special value:
  - DontNeedData suggests not to read data from input. It happens when the operator reads data from grpc streams or the operator was blocked last time and need to digest current blocking record at this time.
  - DontRequireData. Given a union operator, it can read from any inputs without caring about the read order. In this case, we return this status.